On Mon, 21 Mar 2022 22:18:04 -0700 Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > On Mon, Mar 21, 2022 at 10:05 PM Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote: > > > > On Mon, 21 Mar 2022 21:35:55 -0700 > > Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > > > > On Mon, Mar 21, 2022 at 7:36 PM Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote: > > > > > > > > Hi Linus and Alexei, > > > > > > > > At first, sorry about this issue. I missed to Cc'ed to arch maintainers. > > > > > > > > On Mon, 21 Mar 2022 17:31:28 -0700 > > > > Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> wrote: > > > > > > > > > On Mon, Mar 21, 2022 at 4:59 PM Linus Torvalds > > > > > <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > > > > > > > > > > > On Mon, Mar 21, 2022 at 4:11 PM Alexei Starovoitov > > > > > > <alexei.starovoitov@xxxxxxxxx> wrote: > > > > > > > > > > > > > > Did you look at the code? > > > > > > > In particular: > > > > > > > https://lore.kernel.org/bpf/164735286243.1084943.7477055110527046644.stgit@devnote2/ > > > > > > > > > > > > > > it's a copy paste of arch/x86/kernel/kprobes/core.c > > > > > > > > > > > > > > How is it "bad architecture code" ? > > > > > > > > > > > > It's "bad architecture code" because the architecture maintainers have > > > > > > made changes to check ENDBR in the meantime. > > > > > > > > > > > > So it used to be perfectly fine. It's not any longer - and the > > > > > > architecture maintainers were clearly never actually cc'd on the > > > > > > changes, so they didn't find out until much too late. > > > > > > > > Let me retry porting fprobe on top of ENDBR things and confirm with > > > > arch maintainers. > > > > > > Just look at linux-next. > > > objtool warning is the only issue. > > > > Actually, there are conflicts with arm tree and Rust tree too. > > I found I missed the objtool annotation patch on IBT series and fixed it. > > 4 arch patches were reverted. > > > > > > > > > > > > > > Not denying that missing cc was an issue. > > > > > > > > > > We can drop just arch patches: > > > > > rethook: x86: Add rethook x86 implementation > > > > > arm64: rethook: Add arm64 rethook implementation > > > > > powerpc: Add rethook support > > > > > ARM: rethook: Add rethook arm implementation > > > > > > > > > > or everything including Jiri's work on top of it. > > > > > Which would be a massive 27 patches. > > > > > > > > > > We'd prefer the former, of course. > > > > > Later during the merge window we can add a single > > > > > 'rethook: x86' patch that takes endbr into account, > > > > > so that multi-kprobe feature will work on x86. > > > > > For the next merge window we can add other archs. > > > > > Would that work? > > > > > > > > BTW, As far as I can see the ENDBR things, the major issue on fprobe > > > > is that the ftrace'ed ip address will be different from the symbol > > > > address (even) on x86. That must be ensured to work before merge. > > > > Let me check it on Linus's tree at first. > > > > > > That's not an issue. Peter tweaked ftrace logic and fprobe plugs > > > into that. > > > The fprobe/multi-kprobe works fine in linux-next. > > > > Yeah, I think fprobe should work because it uses > > ftrace_location_range(func-entry, func-end) for non-x86 arch. > > > > > > > > bpf selftest for multi kprobe needs this hack: > > > diff --git a/tools/testing/selftests/bpf/progs/kprobe_multi.c > > > b/tools/testing/selftests/bpf/progs/kprobe_multi.c > > > index af27d2c6fce8..530a64e2996a 100644 > > > --- a/tools/testing/selftests/bpf/progs/kprobe_multi.c > > > +++ b/tools/testing/selftests/bpf/progs/kprobe_multi.c > > > @@ -45,7 +45,7 @@ static void kprobe_multi_check(void *ctx, bool is_return) > > > __u64 addr = bpf_get_func_ip(ctx); > > > > > > #define SET(__var, __addr, __cookie) ({ \ > > > - if (((const void *) addr == __addr) && \ > > > + if (((const void *) addr == __addr + 4) && \ > > > (!test_cookie || (cookie == __cookie))) \ > > > > Hmm, this is an ugly hack... You need to use actual ftrace addr, instead of > > symbol addr. With IBT series, you can use ftrace_location(symbol-addr) to > > get the ftrace-addr. (e.g. addr == ftrace_location(__addr) should work) > > It's a temporary hack. > bpf prog cannot call an arbitrary function like ftrace_location. > > > > > > > to pass when both CONFIG_FPROBE=y and CONFIG_X86_KERNEL_IBT=y. > > > The test is too strict. It didn't account for the possibility of endbr. > > > > > > So I'm inclined to drop only 4 arch patches instead of the whole thing. > > > > OK, but it is hard to understand how it works without knowing rethook itself. > > I would like to send whole v13 patch series to arch maintainers. > > fprobe is a glorified kprobe and pretty simple code by itself. > It's too late for v13. Please send 'rethook: x86' patch only > with endbr annotations and get it acked. > The other archs will wait until the next merge window. Hmm, I found the bpf-next tree doesn't have the IBT, so the annotation doesn't work. However, the tip tree doesn't have the rethook itself. If I send the patch to x86 maintainers to be reviewed, I need at least arch independent rethook patch. Or, can you rebase the bpf-next on the tip tree's IBT annotation series? Thank you, -- Masami Hiramatsu <mhiramat@xxxxxxxxxx>
